Application of Data Mining Tools for Classification of Protein Structural Class from Residue Based Averaged NMR Chemical Shifts

Abstract

The number of protein sequences deriving from genome sequencing projects is outpacing our knowledge about the function of these proteins. With the gap between experimentally characterized and uncharacterized proteins continuing to widen, it is necessary to develop new computational methods and tools for protein structural information that is directly related to function. Nuclear magnetic resonance (NMR) provides powerful means to determine three-dimensional structures of proteins in the solution state. However, translation of the NMR spectral parameters to even low-resolution structural information such as protein class requires multiple time consuming steps. In this paper, we present an unorthodox method to predict the protein structural class directly by using the residue’s averaged chemical shifts (ACS) based on machine learning algorithms. Experimental chemical shift information from 1491 proteins obtained from Biological Magnetic Resonance Bank (BMRB) and their respective protein structural classes derived from structural classification of proteins (SCOP) were used to construct a data set with 119 attributes and 5 different classes. Twenty four different classification schemes were evaluated using several performance measures. Overall the residue based ACS values can predict the protein structural classes with 80 % accuracy measured by Matthew Correlation coefficient. Specifically protein classes defined by mixed αβ or small proteins are classified with > 90% correlation. Our results indicate that this NMR-based method can be utilized as a low-resolution tool for protein structural class identification without any prior chemical shift assignments.

I. Introduction

The secondary protein structure was postulated over 60 years ago by Pauling and Corey, who predicted the existence of two local periodic motifs: the α-helix and the β-sheet [1, 2]. By understanding the importance of the relationship between primary and secondary structures of proteins, it can aid in the types of ways that a protein folds. [3–12]. Specifically, the secondary structure is widely used in a number of structural biology applications, such as structure comparison [13], classification [14–16], and visualization [17]. The secondary structures can be used to determine the family, superfamily, and tertiary fold of the underlying protein [18, 19].

Knowledge of the three-dimensional structure of proteins is integral to understanding their functions. A wide range of computational methods are employed to estimate the properties of secondary, tertiary, and quaternary structures proteins. However, experimental methods to provide quantitative information at atomic resolution are limited to NMR spectroscopy and X-ray crystallography. Specifically, NMR spectroscopy, has a proven success at screening large number of proteins in the structural genomics pipeline [20]. However, both NMR and X-ray crystallography approaches are relatively more time and resource consuming procedures in comparison with computational methods. The demands of high throughput proteomics and structural genomics necessitate the development of new, faster experimental methods for providing structural information.

It was first observed in 1957 [21] that nuclear chemical shifts can be powerful indicators of biopolymer structural type. Over the years, chemical shifts have provided detailed information about the nature of hydrogen exchange dynamics, ionization and oxidation states, ring current influence of aromatic residues, and hydrogen bonding interactions [22]. Several review articles describe a wide variety of experimental and computational methods for correlating chemical shifts with protein three-dimensional structure [22–27].

We have demonstrated that the averaged chemical shift (ACS) of a protein backbone nucleus correlates well with both the secondary structure content (SSC) [28, 29] and structural class [30] of the protein. The correlation between the structures can enable the evaluation of the SSC of proteins can aid in the resonance assignment (unique identification of NMR spectral lines with particular nuclear spins within the protein). The ACS values used in these methods are average chemical shifts of all the amino acid residues in the protein. However, there is evidence in the literature that individual amino acids show intrinsic propensities towards certain secondary structure types [7, 31–34]. Furthermore, there is strong correlation between the number of ACS values to the total number of residue types in the given primary sequence to a maximum of 20 naturally occurring amino acids. Extensive research suggests five backbone (¹Hα, ¹³Cα, HN, ¹⁵N and ¹³C) and one side-chain (¹³Cβ ) nuclei are sensitive to changes in the protein structure [27]. Protein structure could be potentially sensitive to 119 ACS values (19 different amino acids and six nuclei and proline with 5 atoms). As the number of proteins with complete chemical shift assignments steadily increasing along with their respective three-dimensional structural information, correlating chemical shifts directly to protein structure is an excellent application for machine learning methods.

In this manuscript we evaluate the application of data mining tools to predict the protein structural class. We have used experimental chemical shift information from BMRB and theoretically estimated values derived from three-dimensional structures. These data mining tools are used to predict protein structural class which validates the general conclusions derived by total ACS. Additionally these tools can provide information regarding the contributions from specific amino acid residues as well as predicting the protein structural class.

II. Materials and Methods

Protein structural information

Structure files were obtained from the Research Collaboratory for Structural Biology (RCSB) (PDB format, http://www.rcsb.org/pdb/) [35]. Since most NMR-STAR files identify several corresponding PDB (protein data bank) structures, it was necessary to examine each entry and choose by inspection the most appropriate PDB ID number. When possible, the PDB ID corresponding to the “best” NMR structure was chosen, though in some cases it was necessary to choose the best X-ray structure (resolution < 2.5 Å). A total of 1491 proteins were found to be suitable, and downloaded from the Protein Data Bank. The secondary structure content (SSC), the total percentage of sheet or helix (α and 3₁₀), was determined using the program PROMOTIF (http://www.biochem.ucl.ac.uk/~gail/promotif/promotif.html) [36].

Protein Chemical Shifts

Protein chemical shift information (NMR-STAR files) obtained from two databases, BioMagResBank (BMRB) (www.bmrb.wisc.edu) [37] and RefDB (www.redpoll.pharmacy.ualberta.ca/RefDB) [38]. BMRB is the first public database to collect chemical shift information from a large number of proteins and RefDB [38] fixes the errors on the files submitted at BMRB ( e.g., reference issues and unassigned or missing resonances). Only proteins with 50 or more amino acid residues, and with at least 70% of their residues assigned chemical shifts, were considered.

In addition to the experimental chemical shifts, the 3D structural information from the respective PDB files were used to estimate the chemical shifts of the proteins using the program Sparta+ [39].

Both the experimental and calculated chemical shifts were then reduced to per-residue averaged chemical shift. The averaged chemical shift (ACS) of a nuclear species “i” was calculated using:

$${AA}_{\mathit{ACS}}^k(i)=\frac{1}{M_k}\sum _{m=1}^{M_k}CS(i,m)$$ [1]

Here i = ¹³CO, ¹³Cα, ¹³Cβ, ¹HN, ¹Hα or ¹⁵N; M_k denotes the total number of residues of type ‘k’ (20 AA’s) with CS values assigned for nucleus species i. CS(i,m) denotes the CS value of the i^ith nucleus at the m^th residue of type ‘k’. If a protein contains all the 20 amino acids then there will be 119 AAACS values.

Protein structural Class

Each protein can be cataloged into one of the six structural classes using SCOP (http://scop.mrc-lmb.cam.ac.uk/scop/): (1) all alpha proteins (α), (2) all beta proteins (β), (3) mixed αβ proteins (αβ), (4) coiled coil, (5) small proteins (s) and membrane/cell surface proteins and multi-domain proteins (m) [40–42]. The distribution of the dataset is shown in Table I.

Table I.

Number of proteins in each structural class.

Protein Structural Class	Abbreviation	Number
All alpha proteins	α	267
All beta proteins	β	317
Alpha and beta proteins (a/b)	αβ	527
Small proteins	s	289
Coiled coil	c	31
Multi-domain proteins	o	60
Total		1491

Data mining

Figure 1 shows the overview of the data mining approach. Six different protein classes defined by the protein structural class represent the classification labels and each protein is represented by 119-dimensional attribute vector ((19 × 6) + (1×5) =119) made of chemical shifts from six different nuclei from BMRB. The number of proteins in the coiled-coil and multi-domain are low compared to other classes (Table I), and were not considered for rest of the analysis. Weka version 3.5.7 (http://www.cs.waikato.ac.nz/ml/weka/), developed by the University of Waikato in New Zealand, is a software collecting a variety of state-of-art machine learning algorithms was employed [43, 44]. 24 different algorithms were performed and the results indicated with a prediction rate higher than 80% in three of the four classes. The list of algorithms are: Bayes Net, Logit Boost, Ridor, NBTree, Multi Class Classifier, Ordinal Class Classifier, SMO, Simple Logistic, END, Random Forest, JRip, Data Near Balanced ND, ND, Decorate, J48, J48 graft, Class Balanced ND, PART, Decision Table, Filtered Classifier, IB1, IBk, Random Sub Space and Kstar. In ten-fold cross-validation, the dataset is split into 10 equal size partitions at random. Each partition is used for testing in turn and the rest is used for training, i.e., each time one-tenth of the dataset is used for testing and the rest for training, and the procedure is repeated 10 times so that each data is used for training and testing exactly once. In this research, 10-fold cross-validation was used to evaluate the classifiers for the basic dataset.

Figure 1.

Overview of the data mining approach to predict protein structural class from the residue based averaged chemical shift values (attributes). DM stands of data mining.

Performance Measures

True positive (TP) provides the measure of number of positive events positive for a virus infection and true negative (TN) provides the number of negative occurrences predicted correctly under a given classification scheme. False positive (FP) gives an estimate of negative events that are incorrectly predicted to be positive, while the false negative (FN) estimated the number of mice that were predicted negative but were positive [45].

For multi-class classification schemes and the sum over rows (i) or columns (j) of the confusion matrix (M) should be considered. For a confusion matrix of dimensionk×k, the TP, TN, FP and FN for the measure (class) ‘n’ could be defined as follows:

$$TP=M_{ii}{|}_{i=n};TN=\sum _{i=1}^k{M}_{ii}{|}_{i\ne n};FP=\sum _{i=1}^k{M}_{ij}{|}_{i\ne n};FN=\sum _{j=1}^k{M}_{ij}{|}_{j\ne n}$$ [2]

These terms were combined to determine the performance of our testing via quantifiable categories such as sensitivity (SN), specificity (SP), positive predictive value (PPV), negative predictive value (NPV), test efficiency/accuracy (TE) and Matthew correlation coefficient (MCC). These quantifiers are defined as follows:

Sensitivity (SN) gives an estimate of the percentage of actual positives identified, while specificity (SP) gives an estimate of the percentage of negatives identified.

$$SN(\%)=\frac{TP\times 100}{TP+FN}$$ [3]

$$SP(\%)=\frac{TN\times 100}{FP+TN}$$ [4]

The effectiveness of a test is evaluated based on two measures namely, positive predictive value (PPV) and negative predictive value (NPV). PPV gives an estimate of the percentage of positive samples that were correctly predicted and NPV gives the percentage of negative samples that were correctly predicted [46, 47].

$$\mathit{PPV}(\%)=\frac{TP\times 100}{TP+FP}$$ [5]

$$\mathit{NPV}(\%)=\frac{TN\times 100}{FN+TN}$$ [6]

The prediction power of a model can be evaluated either by test efficiency (TE) or Mathew correlation coefficient (MCC) [48]. Test efficiency is also referred as test accuracy. The MCC is in essence a correlation coefficient between the observed and predicted classifications; it returns a value between −100% and +100%. A coefficient of +100% represents a perfect prediction, 0% no better than random prediction and −100% indicates total disagreement between prediction and observation [49]. TE and MCC are defined as follows.

$$TE(\%)=\frac{(TP+TN)\times 100}{TP+TN+FP+FN}$$ [7]

$$\mathit{MCC}(\%)=\frac{(TP\times TN-FP\times FN)\times 100}{\sqrt{(TP+FP)(TP+FN)(TN+FN)(TN+FP)}}$$ [8]

The data analyses were performed using scripts written in awk and perl on the Linux workstation. These scripts, as well as a complete list of proteins studied, their BMRB accession numbers, PDB codes, and secondary structure contents, are available from the authors.

III. Results

Figure 2 shows the hierarchical clustering of the residue based averaged chemical shifts between α-helices and β-strands for ¹³Cα (Fig. 2a) and ¹Hα (Fig. 2b). Hierarchical cluster analysis can be used to investigate how the individual chemical shifts of the different nuclei within a particular amino acids contribute to specific structural class of the protein. In this analysis all the proteins were classified using the corresponding amino acid chemical shift values into clusters. Protein secondary structural classes are indicated by dendrograms at the left, and their constituting amino acid chemical shifts (for each nuclei) clusters are indicated by color-coded dendrograms on the left side of the heat map. The extent of variation in the chemical shift values for each protein, is depicted by the intensity scale. The signal level in arbitrary units ranged from “−3 to 3,” with green as the minimum and red as the maximum signal intensities. The relative scaling of the ¹³Cα chemical shifts among the 20 residues indicate the ACS values can be divided into two major groups. Nine amino acid residues Cys, Trp, Arg, Gln, Pro, Phe, His, Met and Tyr fall in one group while the rest in the second group based on the ¹³Cα ACS values, as shown by the dendrograms on the top (Fig. 2). The residue based on ¹Hα ACS values were grouped with the remaining ten amino acids (Met, Cys, Trp, Phe, Tyr, His, Asn, Asp, Pro and Ser) in the first group. This combination of chemical shifts from particular groupings of the amino acids suggest specific combinations towards forming the structural classes.

Figure 2.

Hierarchical clustering of ¹³Cα and ¹Hα residue based ACS values. Natural grouping of the amino acids residues (top) with respect to the experimentally determined residue based ACS values (along the side). Euclidian distance metric was used for clustering. The clustering processes generated natural groupings proteins (secondary structure content) in the form of dendrograms (left color coded) and the respective contributions of the amino acids that constitute the proteins (top dendrograms color coded) with the total changes presented as heat maps. Intensity scales are shown in arbitrary units shown by the scale at the bottom of the panel ranging from green to red in a relative scale. Names of amino acid residues above each heat map represent various combinations and those on the right (color coded) represent either predominantly α helical or β-sheet. The amino acids are grouped into two major clusters based on the profiles. All α and β proteins groups separately as marked for both the ¹³Cα and ¹Hα nuclei.

The residue’s specific average chemical shifts between the α (black symbols) and β (red symbols) class of the proteins, the ¹Hα and ¹³Cα pairs for all the 20 amino acids are plotted in Figure 3. Figure 3 indicates both the experimental and calculated chemical shifts with the overall distributions similar between them. Long side-chain residues (Ile, Leu and Val) differentiate between the helical and strand shifts much more clearly as suggested by the hierarchical clustering (Fig. 2). The distribution of ACS values is narrow for most of the residues except for Cys in both the experimental and calculated values.

Figure 3.

Residue based ACS distribution of the heteronuclear pair ¹³Cα and ¹Hα for each residue type. Experimental (left) and Calculated ACS (right) values for each residue noted by three letter amino acid code on each frame. Chemical shift scaling for all the frames is same except for the Gly residues along the ¹³Cα axis.

Figure 4 shows the experimental (Fig. 4a) and calculated (Fig. 4b) dispersions of ¹HN and ¹⁵N chemical shift dispersions. Residue based ACS values of ¹HN and ¹⁵N show a broader distribution than the corresponding ¹³Cα-¹Hα pairs. Chemical shift dispersions of other nuclei and in particular ¹³C (carbonyl) spins, show a similar distinction between the helical and strand conformational classes (data not shown). Most of the major features, observed in the correlation between ACS values of all residues to protein structural class were observed [30]. Notably increasing the β-sheet characteristics shifts both the ¹³Cα resonance upfield while that of the ¹Hα resonances downfield with an opposite effect between the ¹⁵N and ¹HN pair (downfield shift of ¹⁵N and upfield shift of ¹HN) (Figures 3–4).

Figure 4.

Residue based ACS distribution of the heteronuclear pair ¹⁵N and ¹HN for each residue type. Experimental (left) and Calculated ACS (right) values for each residue noted by three letter amino acid code on each frame. The α and b class proteins are differentiated by black and red symbols, respectively.

Upon including all the structural classes (classification labels) and their qualitative attributes (all 119), these factors contributed to the efficiency of the proteins classification scheme. To determine the best machine learning algorithm for the predictive task, 24 classifiers from WEKA that gave positive results were used using their default parameters. The performance of the algorithms was determined by sensitivity (SN), specificity (SP), positive and negative predictive values (PPV and NPV), test efficiency (TE) and Matthews’s correlation coefficients (MCC) in a 10-fold cross-validation analysis using the experimental or calculated residue based chemical shifts. Of the various classification algorithms, Table II (experimental chemical shifts) and Table III (calculated chemical shifts) present the top performers that classifies all the 4 classes of proteins with at least 80% efficiency. Four algorithms (Bayes Net, Logit Boost, SMO and Random Forest) performed the best classification for the experimental data set, while three algorithms (Bayes Net, Logit Boost and Attribute Selected Classifier) produced good overall results (> 80% MCC) for the calculated chemical shifts. The performance measures of the residue based averaged chemical shifts for the calculated set (Table III) is slightly better than the experimental shifts.

Table II.

Performance of protein structural class prediction using experimental chemical shifts¹

Algorithm	All agr;						All β
	SN	SP	PVP	NPV	TE	MCC	SN	SP	PVP	NPV	TE	MCC
Bayes Net	58.8%	69.7%	69.7%	61.7%	63.0%	86.8%	93.5%	89.6%	88.2%	81.3%	76.5%	78.2%
Logit Boost	76.7%	80.1%	68.5%	25.0%	78.5%	80.9%	94.2%	91.6%	98.2%	92.1%	71.8%	80.1%
SMO	76.1%	78.7%	71.3%	1.7%	48.4%	68.4%	93.9%	91.5%	99.8%	90.0%	60.0%	80.8%
Random Forest	86.0%	74.5%	60.9%	0.0%	52.9%	64.9%	95.5%	94.7%	100.0%	91.0%	60.6%	84.0%
Algorithm	Mixed α/β or α+β						Small Proteins
	SN	SP	PVP	NPV	TE	MCC	SN	SP	PVP	NPV	TE	MCC
Bayes Net	72.7%	23.6%	51.1%	74.3%	90.3%	88.2%	97.5%	87.6%	75.0%	87.6%	83.9%	86.7%
Logit Boost	73.3%	44.1%	75.7%	84.5%	94.2%	89.6%	95.9%	93.2%	79.2%	91.0%	85.7%	94.4%
SMO	76.4%	33.3%	60.1%	82.1%	93.1%	89.2%	94.3%	84.9%	71.3%	90.1%	85.9%	94.1%
Random Forest	81.1%	NA	64.6%	88.0%	92.2%	86.7%	94.3%	86.1%	73.0%	90.4%	85.5%	94.3%

¹SN (sensitivity), SP (Specificity), PVP (Positive predictive value), NPV (Negative predictive value), TE (Test efficiency) and MCC (Mathew Correlation Coefficient).

Table III.

Performance of protein structural class prediction using calculated chemical shifts

Algorithm	All α						All β
	SN	SP	PVP	PVN	TE	MCC	SN	SP	PVP	PVN	TE	MCC
Bayes Net	74.9%	89.3%	83.2%	50.9%	78.1%	92.8%	95.5%	91.7%	96.8%	91.3%	87.0%	84.8%
Logit Boost	81.0%	83.6%	70.2%	36.8%	69.1%	82.9%	95.1%	89.3%	99.2%	92.9%	75.6%	83.3%
Attribute Selected Classifier	77.2%	76.0%	71.2%	24.6%	58.7%	81.6%	93.1%	89.7%	97.2%	87.5%	74.6%	77.1%
Algorithm	Mixed α/β or α+β						Small Proteins
	SN	SP	PVP	PVN	TE	MCC	SN	SP	PVP	PVN	TE	MCC
Bayes Net	76.5%	44.6%	72.3%	85.1%	97.0%	94.4%	97.5%	93.5%	85.8%	94.2%	89.6%	94.6%
Logit Boost	67.8%	70.0%	74.8%	87.0%	95.2%	90.3%	96.7%	90.8%	82.2%	92.5%	84.6%	96.0%
Attribute Selected Classifier	71.0%	33.3%	58.7%	83.6%	92.7%	89.8%	95.8%	87.5%	79.8%	89.1%	84.8%	93.4%

¹SN (sensitivity), SP (Specificity), PVP (Positive predictive value), NPV (Negative predictive value), TE (Test efficiency) and MCC (Mathew Correlation Coefficient).

Protein classes defined by either all β, αβ/α+β or small proteins were well classified, when experimental or calculated chemical shifts were used. In the case of experimental chemical shifts, Logit Boost algorithm is able to classify all the four classes with more than 80% efficiency according to MCC (Table II), with 80.9% for all α, 80.1% for all β, 89.6% for αβ/α+β, and 94.1% for small proteins. For the calculated chemical shifts, Bayes Net produced the best results with 92.8% for all α, 84.8% for all β, 94.4% for αβ/α+β, and 94.6% for small proteins. Residue based ACS values provided a higher correlation (based on MCC) in comparison to linear correlations derived with respect to complete average of chemical shifts. For example, the coefficients of correlation between ACS and sheet content is 0.84 for H^α, and 0.71 for H^N [29].

IV. Discussion and Conclusion

In an effort to explore new methods for the efficient identification of protein structures using NMR, we have investigated the degree to which residue based ACS can be used as a low-resolution structural parameter. The criteria defined to generate the empirical correlations (number of residues > 50, at least 70 % complete chemical shift assignments and 3D structural resolution < 2.5 Å), establish a consistent and improved relationship between four different protein structural classes and residue based ACS values drawn from two separate databanks. Residue based ACS values increased the number of dimensions (attributes) by 20 fold leading to a better discrimination of the classification categories. The results obtained using either experimental chemical shifts directly from the deposited data, or chemical shifts calculated from the 3D structures provides comparable results, with the calculated shifts performing slightly better than experimental values (Table II and III). Figure 5 shows the correlation plot between the calculated and experimental chemical shifts for the nuclei ¹Hα, ¹³Cα and ¹³CO for the all-α and all-β classes. This suggests that the residues based on ACS values are inherently sensitive to the protein structural class.

Figure 5.

Comparison of the experimental (along the Y-axis) and calculated (along X-axis) residue based chemical shifts for helical (top row) and strand conformations (bottom row). Left, middle and the right row represent correspond to ¹Hα, ¹³Cα and ¹³C (carbonyl) nuclei. Each residue is identified by different symbol as noted on the right side of the plot.

Determination of protein structural class, particularly in the absence of chemical shift assignments and primary sequence information could be valuable in a structural proteomics pipeline. One such notable approach was CSSI-PRO presented by Swain and Atreya [34]. Combination of Shifts for secondary structure Identification in Proteins (CSSI-PRO) is based on the detection of specific linear combination of backbone ¹H^α and ¹³C′ chemical shifts in a two-dimensional (2D) NMR experiment. Linear combinations of shifts facilitated editing of residues belonging to α-helical/β-strand regions into distinct spectral regions nearly independent of the amino acid type, thereby allowing the estimation of overall secondary structure content of the protein. In this method a comparison of the predicted vs. experimental secondary structural content for 237 proteins provided a correlation of more than 90% and an overall rmsd of 7.0%. The hierarchical clustering analyses (Fig 2) reflect a similar principle; the profile (combination) of chemical shifts from each residue type contributes differently for the different protein structural classes.

Protein secondary structure prediction algorithms either estimate the backbone dihedral angles (ϕ and ψ) or actual secondary structure definition. The methods such as TALOS (TALOS+) [50, 51], SHIFTOR [52], PREDITOR (+) [53], DANGLE [54] or PROMEGA [55] estimate the dihedral angles while CSI (CSI 2.0) [56], PSSI [57], PsiCSI [58], PLATON [59], PECAN [60] or 2DCSI [61] estimate the secondary structure definition. The program that is relevant for chemical shift prediction includes SHIFTX [62], SPARTA (+) [39], CAMSHIFT [63], SHIFTS [64] or PROSHIFT [65]. In an authoritative review article, Wishart has compared the various prediction approaches (Table 11 [27]). SPARTA+ [39] is one of the top performing program is used here to estimate the chemical shifts from the three dimensional structure of the protein. SPARTA+ Uses an artificial neural network approach and includes a more complete consideration of various structural/dynamic parameters in proteins and able to predict chemical shifts for backbone and ¹³C^β atoms with modestly improved accuracy, compared with other similar chemical shift prediction approaches listed above. SPARTA+ predicted chemical shifts includes structural/dynamic factors, i.e., χ₂ torsion angle, H-bonding and electric fields, as well as an averaging procedure over the outputs from three separated neural networks.

Computational methods often play a primary role in initial predictions of protein structure; specifically in regards to the protein structural class. These methods are typically invoked even before a protein is expressed or extracted for any biophysical characterization. Our results show that residue based ACS values clearly distinguish the four different protein classes, α, β, mixed αβ and small proteins are classified either by SCOP. Multiple algorithms classify the protein structural classes with fairly good efficiency as described by the MCC. The quality of the protein structural class is affected by several known factors that include the size of the protein database, quality of the chemical shift data and classification algorithms. Most of the algorithms presented here show indicate moderate to good performance measured in terms of Mathew correlation coefficient (MCC 80–95%).

Prior to collecting several days’ worth of NMR spectra for structure determination, other biophysical methods are generally adopted to infer secondary structural information about the protein of interest. In particular, circular dichroism (CD) spectroscopy is extensively used to estimate the secondary structure content of medium-sized proteins. However, CD spectrum does not provide information on protein structural class. Furthermore, NMR spectral information has seldom been used to obtain relatively low-resolution structural information, such as protein structural class. In some cases, the results of CD are used to determine whether it is feasible to obtain complete, three-dimensional structural information for a particular protein, using NMR. This suggests the critical importance of evaluating whether data obtained from NMR itself can be used to estimate secondary structure content. Lee and Cao have addressed this question extensively in their comprehensive study [66], and have shown that the correlation between NMR- and CD-based secondary structure estimation is poor. Further, while CD spectroscopy is more suitable for studying relatively small proteins and polypeptides, the characterization of larger molecules requires NMR. Thus making NMR based low resolution approaches complementary to the CD experiments.

It must be emphasized that ACS-based methods do not provide an alternative to conventional NMR-based experiments, and should only be considered initial predictors of protein class or secondary structure content. ACS methods might provide a novel technique for monitoring protein structural changes in real time, such as in protein folding experiments. Such methods might also be used to detect major structural changes that occur upon protein-protein, protein-DNA/RNA, and other complex formations, to provide some direct experimental structural information in situations in which other techniques are incapable of doing so (e.g., in studies of large and/or highly disordered proteins), and to facilitate initial protein fold identification in high throughput proteomics applications.

Highlights.

• Data mining methods to protein chemical shifts to structural class identification

• NMR without chemical shift assignments for protein structural class prediction.

• Averaged chemical shift (ACS) based approach provides 80%–90% accuracy via MCC.